DEER: Disentangled Mixture of Experts with Instance-Adaptive Routing for Generalizable Machine-Generated Text Detection

Researchers have proposed a new framework called DEER, which improves the detection of machine-generated text by decoupling domain-local and domain-invariant knowledge into specialized expert modules.

DEER, or Disentangled Mixture of Experts, employs a reinforcement learning-driven router to select expert pathways based on instance-level detection rewards^[1]. This approach allows DEER to outperform state-of-the-art detectors, achieving average F1 improvements of 1.28% and 2.92%, and accuracy gains of 1.35% and 2.26% on in-domain and out-of-domain datasets, respectively^[1]. Meanwhile, a related study on arxiv.org introduced Selective Sinkhorn Routing (SSR) for Sparse Mixture-of-Experts (SMoE) models, which derives gating scores directly from the transport map and improves training efficiency, accuracy, and robustness to input corruption^[2]. The SSR method addresses the issue of existing SMoE methods often depending on auxiliary objectives or additional trainable components^[2]. By combining these advancements, researchers are making progress in developing more robust and efficient AI models for detecting machine-generated text and improving the scalability of SMoE models.